Appearance
We test local AI with everyday use in mind, comparing task times and output quality across different hardware. Explore the results and samples to find models and configurations that help you work more efficiently. How we test.
NVIDIA RTX 5090 Laptop GPU
24 GB GDDR7 VRAM
896 GB/s memory bandwidth
Apple M5 Max
64 GB unified memory
614 GB/s memory bandwidth
NVIDIA GB10
128 GB unified memory
273 GB/s memory bandwidth
AMD Ryzen AI Max+ 395
128 GB unified memory
256 GB/s memory bandwidth
Results reflect the selected setup for each system. A dash (–) indicates no reported result.
| Model | Benchmark / Metric | Olares One | Mac Studio | NVIDIA DGX Spark | Beelink GTR9 Pro |
|---|---|---|---|---|---|
| Qwen3.8‑27B | IFBench Evaluation time xhigh reasoning, 64k output cap | 8.7 h | 12.2 h | 18.2 h | 33.1 h |
| Qwen3.8‑Flash‑Next | IFBench Evaluation time 300 prompts, xhigh reasoning, 65,536-token output limit | 7.7 h | – | – | – |
| Qwen3.8‑27B | GPQA Diamond Evaluation time xhigh reasoning, 64k output cap | 11.5 h | 14.9 h | – | – |
| Qwen3.8‑27B | Inference speed Token generation, single request 20 prompts, medium reasoning | 93 tok/s | 74 tok/s | 45 tok/s | 28 tok/s |
| Qwen3.8‑Flash‑Next | Inference speed Token generation, single request 20 prompts, medium reasoning, 3,072-token output limit | 166 tok/s | 130 tok/s | 80 tok/s | – |
| Qwen‑Image‑2.1 Base | Text to image Time per image | 12 s | 1.6 min | 27 s | 1.6 min |
| Qwen‑Image‑2.1 Base | Image edit Time per image | 17 s | 1.9 min | 37 s | – |
| Qwen‑Image‑2.1 Turbo | Image edit Time per image | 5.2 s | 30 s | 13 s | 47 s |
| MiniMax H3 FastH3 v2 | Text to video Sampling time nvfp4, 720p, 5 s | 78 s | 11 min | – | – |